Proper orthogonal decomposition

methods for partial differential

equations Chen
Visit to download the full and correct content document:
https://ebookmass.com/product/proper-orthogonal-decomposition-methods-for-partial-
differential-equations-chen/
Proper Orthogonal Decomposition
Methods for Partial Differential
Equations
Mathematics in Science and
Engineering
Proper Orthogonal
Decomposition
Methods for Partial
Differential Equations

Zhendong Luo
School of Mathematics and Physics
North China Electric Power University
Beijing, China

Goong Chen
Department of Mathematics
Texas A&M University
College Station, TX, USA

Series Editor
Goong Chen
Academic Press is an imprint of Elsevier
125 London Wall, London EC2Y 5AS, United Kingdom
525 B Street, Suite 1650, San Diego, CA 92101, United States
50 Hampshire Street, 5th Floor, Cambridge, MA 02139, United States
The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom
Copyright © 2019 Elsevier Inc. All rights reserved.
No part of this publication may be reproduced or transmitted in any form or by any means, electronic or
mechanical, including photocopying, recording, or any information storage and retrieval system, without
permission in writing from the publisher. Details on how to seek permission, further information about
the Publisher’s permissions policies and our arrangements with organizations such as the Copyright
Clearance Center and the Copyright Licensing Agency, can be found at our website:
www.elsevier.com/permissions.
This book and the individual contributions contained in it are protected under copyright by the Publisher
(other than as may be noted herein).
Notices
Knowledge and best practice in this field are constantly changing. As new research and experience
broaden our understanding, changes in research methods, professional practices, or medical treatment
may become necessary.
Practitioners and researchers must always rely on their own experience and knowledge in evaluating and
using any information, methods, compounds, or experiments described herein. In using such information
or methods they should be mindful of their own safety and the safety of others, including parties for
whom they have a professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any
liability for any injury and/or damage to persons or property as a matter of products liability, negligence
or otherwise, or from any use or operation of any methods, products, instructions, or ideas contained in
the material herein.

Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from the Library of Congress

British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library

ISBN: 978-0-12-816798-4

For information on all Academic Press publications

visit our website at https://www.elsevier.com/books-and-journals

Publisher: Candice Janco

Acquisition Editor: Scott J. Bentley
Editorial Project Manager: Devlin Person
Production Project Manager: Maria Bernard
Designer: Matthew Limbert
Typeset by VTeX
Foreword and Introduction

We are living in an era of internet, WiFi, mobile apps, Facebook, Twitters, In-
stagram, selfies, clouds, . . . – surely we have not named them all, but one thing
certain is that these all deal with digital and computer-generated data. There is
a trendy name for the virtual space of all these things together: big data. The
size of the big data space is ever-growing with an exponential rate. Major issues
such as analytics, effective processing, storage, mining, prediction, visualiza-
tion, compression, and encryption/decryption of big data have become problems
of major interest in contemporary technology.
This book aims at treating numerical methods for partial differential equa-
tions (PDEs) in science and engineering. Applied mathematicians, scientists,
and engineers are always dealing with big data, all the time. Where do their big
data originate? They come, mostly, from problems and solutions of equations of
physical and technological systems. A large number of the modeling equations
are PDEs. Therefore, effective methods and algorithms for processing and re-
solving such data are much in demand. This is not a treatise on general big data.
However, the main objective is to develop a methodology that can effectively
help resolve the challenges of dealing with large data sets and with speedup in
treating time-dependent PDEs.
Our approach here is not the way in which standard textbooks on numerical
PDEs are written. The central theme of this book is actually on the technical
treatment for effective methods that can generate numerical solutions for time-
dependent PDEs involving only a small set of data, but yield decent solutions
that are accurate and suitable for applications. It reduces data storage, CPU time,
and, especially, computational complexity – several orders of magnitude. The
key idea and methodology is proper orthogonal decomposition (POD), from
properties of eigensolutions to a problem involving a large data set. (Indeed,
POD has been known as an effective method for big data even before the term
big data was coined.) In the process, we have developed the necessary mathe-
matical methods and techniques adapting POD to fundamental numerical PDE
methods of finite differences, finite elements, and finite volume in connection
with various numerical schemes for a wide class of time-dependent PDEs as
showcases.

xi
xii Foreword and Introduction

PDES AND THEIR NUMERICAL SOLUTIONS

Physical, biological, and engineering processes are commonly described by
PDEs. Such processes are naturally dynamic, meaning that their time evolution
constitutes the main features, properties, and significance of the system under
investigation or observation. The spatial domains of definition of the PDEs are
usually multidimensional and irregular in shape. Thus, there are fundamental
difficulties involving geometry and dimensionality. The PDEs themselves can
also take rather complex forms, involving a large variety of nonlinearities, sys-
tem couplings, and source terms. These are inherent difficulties of the PDEs
that compound those due to geometry and dimensionality. In general, exact (an-
alytic) nontrivial solutions to PDEs are rarely available. Numerical methods and
algorithms must be developed and then implemented on computers to render ap-
proximate numerical solutions. Therefore, computations become the only way
to treat PDE problems quantitatively. The study of numerical solutions for PDEs
is now a major field in computational science that includes computational math-
ematics, physics, engineering, chemistry, biology, atmospheric, geophysical and
ocean sciences, etc.
Computational PDEs represent an active, prosperous field. New methods
and developments are constantly emerging. However, three canonical schemes
stand out: finite difference (FD), finite element (FE), and finite volume element
(FVE) methods. These methods all require the division of the computational do-
main into meshes. Thus, they involve many degrees of freedom, i.e., unknowns,
which are related to the number of nodes of mesh partition of the computa-
tional domains. For a routine, real-world problem in engineering, the number of
unknowns can easily reach hundreds of thousands or even millions. Thus, the
amount of computational load and complexity is extremely high. The accuracy
of numerical solutions is also affected as truncation errors tend to accumulate.
For a large-scale problem, the CPU time on a supercomputer may require days,
weeks, or even months. It is possible, for example, if we use these canonical
methods of FD, FE, and FVE to simulate the weather forecast in atmospheric
science, after a protracted period of computer calculations, that the output nu-
merical results have already lost their significance as the days of interest are
bygone.
There are two ways of thinking for the resolution of these difficulties. First,
one can think of computer speedup by building the best supercomputers with
continuous refinement. As of June 2016, the world’s fastest supercomputer
on the TOP500 (http://top500.org) supercomputer list is the Sunway Taihu-
Light in China, with a LINPACK benchmark score of 93 PFLOPS (Peta, or
1015 , FLOPS), exceeding the previous record holder, Tianhe-2, by around 59
PFLOPS. Tianhe-2 had its peak electric power consumption at 17.8 MW, and
its annual electricity bill is more than $14 million or 100 million Chinese Yuan.
Thus, most tier-1 universities cannot afford to pay such a high expense. The
second option is to instead develop highly effective computational methods that
can reduce the degrees of freedom for the canonical FD, FE, and FVE schemes,
 Foreword and Introduction xiii

lighten the computational load, and reduce the running CPU time and the accu-
mulation of truncation errors in the computational processes. This approach is
based on cost-optimal, rational, and mathematical thinking and will be the one
taken by us here.
The focal topic of the book, the POD method (see [56,60]), is one of the
most effective methods that aims exactly at helping computational PDEs.

THE ADVANTAGES AND BENEFITS OF POD

Reduce the degrees of freedom of numerical computational models for time-
dependent PDEs, alleviate the calculation load, reduce the accumulation of
truncated errors in the computational process, and save CPU computing time
and resources for large-scale scientific computing.

POD in a Nutshell
The POD method essentially provides an orthogonal basis for representing a
given set of data in a certain least-squares optimal sense, i.e., it offers ways to
find optimal lower-dimensional approximations for the given data set.

A Brief Prior History of the Development of POD

The POD method has a long history. The predecessor of the POD method was
an eigenvector analysis method, which was initially presented by K. Pearson in
1901 and was used to extract targeted, main ingredients of huge amounts of data
(see [132]). (The trendy name of such data is “big data”.) Pearson’s data mining,
sample analysis, and data processing techniques are relevant even today.
However, the method of snapshots for POD was first presented by Sirovich
in 1987 (see [150]). The POD method has been widely and successfully ap-
plied to numerous fields, including signal analysis and pattern recognition (see
[43]), statistics (see [60]), geophysical fluid dynamics or meteorology (see
also [60], [60] or [78]), and biomedical engineering (see [48]). For a long
time since 1987, the POD method was mainly used to perform the princi-
pal component analysis in statistical computations and to search for certain
major behavior of dynamic systems (see reduced-order Galerkin methods for
PDEs, proposed in the excellent work in 2001 by Kunisch and Volkweind [62,
63]). From that moment forth, the model reduction or reduced basis of the
numerical computational methods based on POD for PDEs underwent some
rapid development, providing improved efficiency for finding numerical so-
lutions to PDEs (see [2,11,15,19,45,48,54,57,58,135,137,138,141,166,167,170,
171,184–186,199]). At first, Kunisch–Volkweind’s POD-based reduced-order
Galerkin methods were applied to reduced-order models of numerical solutions
for PDEs with error estimates presented in [62,63]. Those error estimates con-
sist of some uncertain matrix norms. In particular, they took all the numerical
solutions of the classical Galerkin method on the total time span [0, T ] in the
xiv Foreword and Introduction

formulation of the POD basis, used them to establish the POD-based reduced-
order models, and then recomputed the numerical solutions on the same time
span [0, T ]. This produces some repetitive computation but not much extra gain.
One begins to ponder how to improve this and, furthermore, how to generalize
the methodology initiated by Kunisch–Volkweind’s work beyond the Galerkin
FE method to other FE methods and also to FD and FVE schemes. This book
aims exactly at answering these questions.

Development of POD for Time-Dependent PDEs

The first author, Zhendong Luo, was attracted to the study of reduced-order nu-
merical methods based on POD for PDEs at the beginning of 2003. At that time,
few or no comprehensive accounts existed and only fragmentary introductions
about POD were available. He spent three years (2003–2005) studying the un-
derlying optimization methods, statistical principles, and numerical solutions
for POD. Then, in 2006, he and his collaborators published their first two pa-
pers for POD methods (see [26,27]). These dealt with oceanic models and data
assimilation.
Afterwards, Luo and his coauthors have established some POD-based
reduced-order FD schemes (see [5,38,40,91,113,118,122,155]) as well as FE
formulations (see [37,39,70,88–90,92,93,100,101,103,109,112,119,123,124,
164]). They deduced the error estimates for POD-based reduced-order solutions
for PDEs of various types since 2007 in a series of papers. They also proposed
some POD-based reduced-order formulations and relevant error estimates for
POD-based reduced-order FVE solutions (see [71,104,106,108,120]) for PDEs
in another series of papers beginning in 2011. These POD-based reduced-order
methods were specific to the classical FD schemes, FE methods, and FVE meth-
ods for the construction of the reduced-order models, in which they extracted
one from every ten classical numerical solutions as snapshots, significantly dif-
ferent from Kunisch–Volkweind’s methods, in which numerical solutions from
the classical Galerkin method were extracted at all instants on the total time
span [0, T ]. Therefore, these POD-based reduced-order methods constitute im-
provements, generalizations, and extensions for Kunisch–Volkweind’s methods
in [62,63]. The reduced-order methods in the above cited work need only repeat
part of the computations on the same time span [0, T ].
Since 2012, Luo and his collaborators have established the following three
main methods:
i. PODROEFD: POD-based reduced-order extrapolation FD schemes (see [6,
7,79,81,94–96,102,110,111,117,121,127,154,158]);
ii. PODROEFE: POD-based reduced-order extrapolation FE methods (see [69,
75,82,83,97,116,159–161,165,179]);
iii. PODROEFVE: POD-based reduced-order extrapolation FVE methods (see
[84,85,87,98,99,114,115,162,163]).
 Foreword and Introduction xv

These POD-based reduced-order extrapolation methods need only adopt the

standard numerical solutions on some initial rather short time span [0, t0 ]
(t0  T ) of, respectively, the classical FD, FE, and FVE schemes as snapshots in
order to formulate the POD basis. Therefore, they have significantly improved
the previous, existing version of the reduced-order models. They do not have
to repeat wholesale computations. The physical significance is that one can use
existing data to forecast the future evolution of nature. Furthermore, our PO-
DROEFD, PODROEFE, and PODROEFVE methods can be treated in a similar
way as the classical FD, FE, and FVE methods, leading to error estimates with
concrete orders of convergence. The application of these POD-based extrapola-
tion methods will provide anyone with the advantages and benefits of POD
mentioned earlier.
The second author, Goong Chen, has strong interests in the computation
of numerical solutions of PDEs arising from real-world applications. He has
constantly been faced with the challenges to deal with the needs for large
data storage, process speedup, effective reduction of order, and the extraction
of prominent physical features from the supercomputer numerical solution of
PDEs. When he noticed that Zhendong Luo had already done significant work
on the POD methods for time-dependent PDEs fitting many of his needs, he
got very excited and proposed that a book project be prepared to publish and
promote this very important topic. This started the collaboration of the authors,
with this book as the outcome. Our collaboration is ongoing, hoping more re-
search papers will be produced in the near future demonstrating the advantages
of POD-based methods. However, G. Chen happily acknowledges that all tech-
nical contributions in this book are to be credited to the first author alone. He
has learned tremendously from the collaboration – this by itself makes the book
project worthwhile and satisfying as far as the second author is concerned.

ORGANIZATION OF THE BOOK

In this book, we aim to provide the technical details of the construction, theoreti-
cal analysis, and implementations of algorithms and examples for PODROEFD,
PODROEFE, and PODROEFVE methods for a broad class of dynamic PDEs.
It is organized into the following four chapters.
Chapter 1 includes four sections. In the first section, we review the basic the-
ory of classical FD schemes. It is intended to ensure the self-containedness of the
book. Then, in the subsequent three sections, we introduce the construction, the
theoretical analysis, and the implementations of algorithms for the PODROEFD
schemes for the following two-dimensional (2D) PDEs: the parabolic equation,
the nonstationary Stokes equation, and the shallow water equation with sedi-
ment concentration, respectively. Examples and discussions are also given for
each equation.
Chapter 2 is similarly structured as Chapter 1, with four sections. There
we begin by reviewing the basic theory of Sobolev spaces and elliptic theory,
xvi Foreword and Introduction

the classical FE method, and the mixed FE (MFE) method. Then we describe
the construction, theoretical analysis, and implementations of algorithms for the
PODROEFE methods for the following 2D PDEs: the viscoelastic wave equa-
tion, the Burgers equation, and the nonstationary parabolized Navier–Stokes
equation (for which the stabilized Crank–Nicolson extrapolation scheme is
used), respectively. Numerical examples and graphics are again illustrated.
Chapter 3 contains three sections, aiming at the treatment of PODROEFVE.
We first introduce the basics of FVE. Then three sections for the construction,
theoretical error analysis, and the implementations of algorithms for the PO-
DROEFVE methods for the following three 2D dynamic PDEs are studied: the
hyperbolic equation, Sobolev equation, and incompressible Boussinesq equa-
tion, respectively, are developed, with concrete examples and illustrations.
Numerical results on these model equations as presented in the book have
demonstrated the effectiveness and accuracy of our POD methods.
Finally, Chapter 4 is a short epilogue and outlook, consisting of concluding
remarks and forward-looking statements.
The book is written to be as self-contained as possible. Readers and students
need only to have an undergraduate level in applied and numerical mathemat-
ics for the understanding of this book. Many parts can be used as a standard
graduate-level textbook on numerical PDEs. The theory, methods, and com-
putational algorithms will be valuable to students and practitioners in science,
engineering, and technology.

ACKNOWLEDGMENTS
The authors thank all collaborators, colleagues, and institutions that have gen-
erously supported our work. In particular, the authors are delighted to acknowl-
edge the partial financial support over the years by the National Natural Science
Foundation of China (under grant #11671106), the Qatar National Research
Fund (under grant #NPRP 8-028-1-001), the North China Electric Power Uni-
versity, and the Texas A&M University.

Zhendong Luo
Beijing, China
Goong Chen
College Station, TX, USA
Chapter 1

Reduced-Order Extrapolation
Finite Difference Schemes
Based on Proper Orthogonal
Decomposition
The key objective of this book is to develop the numerical treatments of
proper orthogonal decomposition (POD) for partial differential equations
(PDEs). With regard to numerical methods for PDEs, the finite difference
(FD) method essentially constitutes the basis of all numerical methods for
PDEs. In order to introduce how POD works, it is natural to start with the
FD method.
For the sake of proper self-containedness, in this chapter we first review the
basic theory of classical FD schemes. We then introduce the construction, theo-
retical analysis, and implementations of algorithms for the POD-based reduced-
order extrapolation FD (PODROEFD) schemes for the two-dimensional (2D)
parabolic equation, 2D nonstationary Stokes equation, and 2D shallow wa-
ter equation with sediment concentration. Finally, we provide some numerical
examples to show what the PODROEFD schemes have over the classical FD
schemes. Moreover, it is shown that the PODROEFD schemes are reliable and
effective for solving above-mentioned PDEs.
The numerical models treated here include both simple equations and cou-
pled systems. The systematic approach we take in this chapter, namely, fol-
lowing the logical sequence of rudiments, modeling equations, error estimates-
stability-convergence, POD methods, error estimates for POD solutions, and
finally concrete numerical examples, will be the standard for all chapters.

1.1 REVIEW OF CLASSICAL BASIC FINITE DIFFERENCE

THEORY
1.1.1 Approximation of Derivative
The FD schemes use difference quotients to approximate derivatives. Denote
uni,j = u(ix, j y, nt) = u(xi , yj , tn ). Then uni±1,j ±1 = u(xi ± x, yj ±
y, tn ), un±1
i,j = u(xi , yj , tn ± t).
Derivative approximations have usually the following four forms.

Proper Orthogonal Decomposition Methods for Partial Differential Equations 1

https://doi.org/10.1016/B978-0-12-816798-4.00006-1
Copyright © 2019 Elsevier Inc. All rights reserved.
2 Proper Orthogonal Decomposition Methods for Partial Differential Equations

1. Approximation to first-order derivative by a forward difference

We have
 n
∂u u(xi + x, yj , tn ) − u(xi , yj , tn )
= lim
∂x i,j x→0 x
u(xi + x, yj , tn ) − u(xi , yj , tn )
= + O(x)
x
uni+1,j − uni,j
= + O(x)
x
uni+1,j − uni,j
≈ . (1.1.1)
x

2. Approximation to first-order derivative by a backward difference

We have
 n
∂u u(xi , yj , tn ) − u(xi − x, yj , tn )
= lim
∂x i,j x→0 x
u(xi , yj , tn ) − u(xi − x, yj , tn )
= + O(x)
x
uni,j − uni−1,j
= + O(x)
x
ui,j − ui−1,j
n n
≈ . (1.1.2)
x

3. Approximation to first-order derivative by a central difference

We have
 n
∂u uni+1,j − uni−1,j
= + O(x 2 )
∂x i,j 2x
uni+1,j − uni−1,j
≈ . (1.1.3)
2x

4. Approximation to second derivative by a second-order central differ-

ence
We have
 n
∂ 2u uni+1,j − 2uni,j + uni−1,j
= + O(x 2 )
∂x 2 i,j x 2
uni+1,j − 2uni,j + uni−1,j
≈ . (1.1.4)
x 2
 Reduced-Order Extrapolation Finite Difference Schemes Chapter | 1 3

1.1.2 Difference Operators

1. The definitions of difference operators
We denote operators I, Ex , Ex−1 , x , ∇x , δx , μx by the following:
I unj = unj ; I is known as the unit operator;
Ex unj = unj+1 ; Ex is known as the forward shift operator and denoted by
Ex = Ex+1 ;
Ex−1 unj = unj−1 ; Ex−1 is known as the backward shift operator and denoted by
Ex− = Ex−1 ;
x unj = unj+1 − unj ; x is known as the forward difference operator and sat-
isfies x = Ex − I ;
∇x unj = unj − unj−1 ; ∇x is known as the backward difference operator and
satisfies ∇x = I − Ex− ;
δx unj = unj+ 1 − unj− 1 ; δx is known as the one step central difference and
2 2
1
−1
satisfies δx = Ex − Ex 2 ;
2

1
μx unj = (uj + 1 + uj − 1 ); μx is known as the average operator and satisfies
2 2 2
1 12 − 12
μx = (Ex + Ex ).
2
2. Composite difference operators
We have
1 1
i. (μδ)x = (Ex − Ex−1 ) = (x + ∇x );
2 2
1
− 12 2
ii. (δx ) = δx δx = (Ex − Ex ) = (Ex − 2I + Ex−1 );
2 2

iii. (δx )n = δx (δxn−1 ); nx = x n−1x = · · · ; ∇xn = ∇x · ∇xn−1 = · · · .

3. Derivative relations with difference operators
We have
 n
∂u x unj
i. = + O(x)—forward difference
∂x j x
∇x unj
= + O(x)—backward difference
x
δx unj
= + O(x 2 )—central difference;
x
 n
∂ 2u δx2 unj
ii. = + O(x 2 )—the second-order central difference
∂x 2 j x 2
2x unj
= + O(x 2 )—the second-order forward difference
x 2
∇x2 unj
= + O(x 2 )—the second-order backward difference.
x 2
4 Proper Orthogonal Decomposition Methods for Partial Differential Equations

1.1.3 The Formation of Difference Equations

1. Explicit FD schemes
The explicit FD scheme implies that the time farthest point values appear
only once. For example,

un+1
j − unj δ
= (un − 2unj + unj−1 )
t x 2 j +1

∂u ∂ 2u
is an explicit FD scheme for = δ 2 , where δ is a positive “diffusion coef-
∂t ∂x
ficient”.
2. Implicit FD schemes
The implicit FD schemes imply that the time farthest point values in the FD
scheme appear more than once. For example,

un+1
j − unj δ
= (un+1 − 2un+1 + un+1
j −1 )
t x 2 j +1 j

∂u ∂ 2u
is an implicit FD scheme for =δ 2.
∂t ∂x
3. Semidiscretized difference schemes
The semidiscretized difference scheme implies that only the spatial variable
is discretized and the time variable is not discretized. For example,
 
du δ
= (uj +1 − 2uj + uj −1 )
dt j x 2

∂u ∂ 2u
is a semidiscretized difference scheme for =δ 2.
∂t ∂x

1.1.4 The Effectiveness of Finite Difference Schemes

1. The error of an FD scheme
For the equation
∂u ∂ 2u
=δ 2,
∂t ∂x
consider an FD scheme discretized by forward difference for time and central
difference for space (FTCS) as follows:

un+1
j − unj δ
= (un − 2unj + unj−1 ),
t x 2 j +1
 Reduced-Order Extrapolation Finite Difference Schemes Chapter | 1 5

which is expanded at reference point (xj , tn ) by Taylor’s formula into

 n   n  n 
∂u ∂ 2u t ∂ 2u x 2 ∂ 4u
−δ 2 + −δ + · · · = 0.
∂t ∂x j 2 ∂t 2 j 12 ∂x 4 j

The above equation is known as the modified PDE, where the first parenthesis
above is called the source equation (i.e., PDE) and the bracket above is called
the remainder (R) or truncation error (TE), denoted by R = TE.
The TE is equal to the source equation (PDE) subtracting the FD equation
(FDE), i.e., TE = PDE − FDE.
The discretization error (DE, i.e., the error of the FD scheme) is equal to
the TE plus the boundary error (BE), i.e., DE = TE + BE. Thus, DE = BE +
PDE − FDE.
The round-off error (ROE) denotes the rounding error of the computing pro-
cedure by the computer and the total error is denoted by CE. Then

CE = DE + ROE = BE + PDE − FDE + ROE.

However, ROE and BE are usually omitted. The error of FD schemes mainly
considers TE, which is directly obtained from the approximation of the deriva-
tive (1.1.1)–(1.1.4).
2. The consistency of an FD scheme

Definition 1.1.1. (1) Let the PDE

ut = Lu (1.1.5)

be discretized by an FD scheme as follows:

 
αμ un+1
j +μ = βγ unj+γ . (1.1.6)
μ γ

When an FD grid is indefinitely refined and if R = T E satisfies the property

that it tends to zero, i.e.,

lim R = lim T E = 0, (1.1.7)

x→0 x→0
t→0 t→0

then the FD scheme is said to be compatible with the source equation.

t
(2) If t = o(x γ ) (γ > 0), i.e., lim = 0, we have R → 0, then the
x→0 x γ
t→0
FD scheme is said to be compatible with the source equation on the conditions
t = o(x γ ) (γ > 0).

3. The stability of an FD scheme

6 Proper Orthogonal Decomposition Methods for Partial Differential Equations

Definition 1.1.2. i. If an error disturbance εjn = ũnj − unj is added at certain time
level t = tn , i.e., ũnj = unj + εjn , and the errors εjn+1 = ũn+1
j − un+1
j of solutions
ũn+1
j and un+1
j obtained from the FD scheme (1.1.6) do not produce extra-large
overall growth, i.e., there is a constant K > 0 independent of n and j such that

εjn+1   Kεjn , or omitting j, εn+1   Kε n ,

where  ·  represents a norm, then when 0 < K < 1, the FD scheme (1.1.6) is
said to be strongly stable, else (when K  1) the FD scheme (1.1.6) is said to be
weakly stable.
ii. If there is no restriction between the time step and the spatial step in the
FD scheme (1.1.6), it is said to be absolutely stable or unconditionally stable.
In this case, the time step t can take a larger size.
iii. If the stability of the FD scheme (1.1.6) is restricted by some relationship
of time step and spatial steps (usually constrained in the form of some inequal-
ities, for example, t = o(x γ ) (γ > 0)), then it is said to be conditionally
stable.
We have the following criteria for the stability of the FD scheme (1.1.6) (see
[192,194]).
Theorem 1.1.1. The FD scheme (1.1.6) is stable if and only if there are two
positive constants t0 and x0 as well as a nonnegative constant K > 0 inde-
pendent of n and j such that the solutions of FD scheme (1.1.6) satisfy

unj   K, n = 1, 2, . . . .

Theorem 1.1.2. If the FD scheme (1.1.6) of PDE (1.1.5) satisfies

un+1
j 
un+1 · unj > 0, < 1, n = 1, 2, · · · , (1.1.8)
j
unj 

where  ·  is a discrete norm, then the FD scheme (1.1.6) is stable.

4. Equivalence between stability and convergency of FD schemes
Definition 1.1.3. i. Let u∗ (x, t) be an exact solution for PDE (1.1.5) and
unj the approximate solutions of the FD scheme (1.1.6) compatible with the
PDE (1.1.5). If when t → 0, x → 0 (i.e., grid is infinitely refined), for any
sequence (xj , tn ) → (x ∗ , t ∗ ) ∈ , we have

lim unj = u∗ (x ∗ , t ∗ ), (1.1.9)

x,t→0

then the FD scheme (1.1.6) is said to be convergent.

The stability and convergence of FD schemes are their intrinsic properties
with the following important equivalence (see [192] or [194, Theorem 1.3.18]).
 Reduced-Order Extrapolation Finite Difference Schemes Chapter | 1 7

Theorem 1.1.3. If the PDE (1.1.5) is well posed, i.e., it has a unique solution
and is continuously dependent on initial boundary value conditions and com-
patible with its FD scheme (1.1.6), then the stability of the FD scheme (1.1.6) is
equivalent to its convergence.

Theorem 1.1.3 is briefly described as “if a PDE is well posed and compatible
with its FD scheme, then the stability of the FD scheme (1.1.6) is equivalent to
its convergence”.

Remark 1.1.1. For linear PDEs, if their FD schemes are stable, then their con-
vergence is ensured. Therefore, it is only necessary to discuss their stability.
However, for nonlinear PDEs, because of the complexity, at present we can
only investigate nearly similar convergence and local stability analysis instead
of global analysis of convergence.

5. Von Neumann’s stability analysis of FD schemes

Von Neumann’s stability analysis of FD schemes is also known as the
Fourier analysis method. In order to analyze the stability of FD schemes, we
first discuss the stability of exact solutions.
i. Stability analysis of exact solution
The initial value problem
⎧
⎨ ∂ ∂
ut = L( , 2 , · · · )u, x ∈ R,
∂x ∂x (1.1.10)
⎩
u(x, 0) = φ(x)

is said to be well posed, if it has a unique and stable solution. A so-called stable
solution means that the solution is continuously dependent on the initial value
and can maintain the boundedness of small disturbances.
Assume that the initial value φ(x) is periodic and expandable into the fol-
lowing Fourier series:

φ(x) = fk eikx , (1.1.11)
k

where the fk s are the Fourier coefficients. Assume that the exact solution u(x, t)
of the initial value problem (1.1.10) is also expandable into the following Fourier
series:

u(x, t) = Fk (t)eikx , (1.1.12)
k

where the Fk (t)s are the Fourier coefficients of the series depending on t.
8 Proper Orthogonal Decomposition Methods for Partial Differential Equations

By substituting (1.1.12) into the first equation of (1.1.10), we obtain

 ∂ ∂ 
Fk (t)eikx = L( , 2,···) Fk (t)eikx
∂x ∂x
k k
 ∂ ∂
= L( , 2 , · · · )eikx · Fk (t). (1.1.13)
∂x ∂x
k

If L is a linear differential operator, we can rewrite (1.1.13) as follows:

 
Fk (t)eikx = L(ik, (ik)2 , · · · )eikx · Fk (t). (1.1.14)
k k

By comparing the coefficients eikx of the LHS and those of the RHS in (1.1.14)
and setting them equal, we have

Fk (t) = L(ik, (ik)2 , · · · ) · Fk (t). (1.1.15)

Eq. (1.1.15) has a general solution as follows:

2 ,··· )t
Fk (t) = ck eL(ik,(ik) . (1.1.16)

By using the initial condition u(x, 0) = φ(x) of (1.1.10), we obtain

fk eikx = Fk (0)eikx , which implies ck = Fk (0) = fk . Thus, (1.1.12) can
k k
be rewritten as follows:
 2 ,··· )t
u(x, t) = fk eL(ik,(ik) · eikx . (1.1.17)
k

Note that the exact solution u(x, t) is said to be stable, if there is a nonnegative
constant M, independent of u, t, and x, such that

u(x, t)L2  Mu(x, 0)L2 . (1.1.18)

Because {eikx } is a set of standard orthogonal bases in L2 (−π, π), we have

 2 ,··· )t
u(x, t)2L2 = | fk eL(ik,(ik) |2
k
 2 ,··· )t
 | fk | sup | eL(ik,(ik)
2
|2 , (1.1.19)
k

u(x, 0)2L2 = | fk |2 , (1.1.20)
k

π 1/2
where u(x, t)L2 = −π | u(x, t) |2 dx is the norm of u(·, t) in L2 (−π, π).
 Reduced-Order Extrapolation Finite Difference Schemes Chapter | 1 9

From (1.1.19)–(1.1.20), we know that

2 ,··· )t
| eL(ik,(ik) | M (1.1.21)

holds if and only if

 2 ,··· )t
u(x, t)2L2 = | fk eL(ik,(ik) |2
k

M 2
| fk |2  M 2 u(x, 0)2L2 . (1.1.22)
k

Thus, the exact solution u(x, t) of (1.1.10) is stable if and only if there is a
nonnegative constant M independent of u, t, and x, such that | eL(ik)t | M.
ii. Stability analysis of an FD scheme
In the following, we analyze the von Neumann stability conditions of an FD
scheme.
Let (1.1.10) have the following FD scheme:
 
αμ un+1
j +μ = βγ unj+γ . (1.1.23)
μ γ

Let the initial time level (n = 0) be denoted as

 
u0j = φ(xj ) = gk0 eikxj = gk0 eikj x . (1.1.24)
k k

Let the nth time level solution unj be denoted by

 
unj = gkn eikxj = gkn eikj x . (1.1.25)
k k

By inserting (1.1.25) into (1.1.23), and by the standard orthogonality of

{eikx }, we obtain

gkn+1 = Ggkn , k ∈ Z, (1.1.26)

where Z is the set of all integers and G is known as the growth factor, which is
denoted by the following equation:
 −1 ⎛ ⎞
 
G= αμ eikμx ·⎝ βγ eikγ x ⎠ . (1.1.27)
μ γ

Let g = (· · · , g−k , · · · , g−2 , g−1 , g0 , g1 , g2 , · · · , gk , · · · ). We have

g n = Gg n−1 , n = 1, 2, · · · . (1.1.28)
10 Proper Orthogonal Decomposition Methods for Partial Differential Equations

By using Eq. (1.1.28), we have

g n 2 = Gg n−1 2 = G2 g n−2 2 = · · · = Gn g 0 2 , n = 1, 2, · · · , (1.1.29)

where  · 2 represents the norm in l 2 . Thus, by (1.1.24) and (1.1.25), from

(1.1.29), we obtain

unj 2  Gn u0j 2 = Gn φ(xj )2 , n = 1, 2, · · · . (1.1.30)

Then, by Theorem 1.1.1, we obtain the following result.

Theorem 1.1.4. The FD scheme (1.1.13) is stable if and only if there are two
positive constants t0 and x0 as well as a nonnegative constant K > 0 inde-
pendent of n and j , such that its growth factor G satisfies

Gn   K, n = 1, 2, · · · . (1.1.31)

Corollary 1.1.5. The FD scheme (1.1.13) is stable if and only if its growth
factor G satisfies

G  1 + O(t), n = 1, 2, · · · . (1.1.32)

iii. Von Neumann’s stability analysis of FD schemes

The condition of (1.1.32) in Corollary 1.1.5 is usually known as the von
Neumann stability condition. It follows that in order to distinguish the stability
of the FD scheme (1.1.13), it is necessary to compute its growth factor G by
formula (1.1.27) and determine the t and x such that (1.1.32) or G  1 is
satisfied.
For specific FD schemes, it is easy to compute their growth factor G by
(1.1.27). By the standard orthogonality of {eikx }, it is necessary to substitute

unj = Gn eikxj , n = 1, 2, · · · (1.1.33)

into the FD scheme (1.1.13), and then, eliminating some common factors, one
can obtain the growth factor G of (1.1.27).
For an FD scheme for two-dimensional linear PDEs, we need only to substi-
tute unj,m = Gn eikxj eikym (n = 1, 2, · · · ) into the FD scheme and then simplify,
so we can also obtain the growth factor G.
Some relative examples can be found in [192].

1.2 A POD-BASED REDUCED-ORDER EXTRAPOLATION

FINITE DIFFERENCE SCHEME FOR THE 2D PARABOLIC
EQUATION
In this section, we introduce the PODROEFD scheme for the two-dimensional
(2D) parabolic equation. The work is based on Luo et al. [111].
 Reduced-Order Extrapolation Finite Difference Schemes Chapter | 1 11

For convenience and without loss of generality, let = (ax , bx ) × (cy , dy )

and consider the following 2D parabolic equation.
Find u such that
⎧
⎪
⎨ ut − u = f, (x, y, t) ∈ × (0, T ),
u(x, y, t) = g(x, y, t), (x, y, t) ∈ ∂ × [0, T ), (1.2.1)
⎪
⎩ u(x, y, 0) = s(x, y), (x, y) ∈ ,

where f (x, y, t), g(x, y, t), and s(x, y) are given source item, boundary func-
tion, and initial value function, respectively, and T is the total time duration.
The main motivation and physical background of the parabolic equation are
the modeling of heat conduction and diffusion phenomena.

1.2.1 A Classical Finite Difference Scheme for the 2D Parabolic

Equation
Let x and y be the spatial steps in the x and y directions, respectively, t the
time step, and unj,k the function value of u at points (xj , yk , tn ) (0  j  J =
[(bx − ax )/x], 0  k  K = [(dy − cy )/y], 0  n  N = [T /t], where
[(bx − ax )/x], [(dy − cy )/y], and [T /t] denote the integral parts of (bx −
ax )/x, (dy − cy )/y, and T /t, respectively).
Thus, the forward difference explicit scheme for the 2D parabolic equa-
tion (1.2.1) at reference point (xj , yk , tn ) is given by
t n
un+1
j,k = uj,k +
n
(u − 2unj,k + unj−1,k )
x 2 j +1,k
t n
+ (u − 2unj,k + unj,k−1 ) + tfj,k
n
. (1.2.2)
y 2 j,k+1
For the FD scheme (1.2.2), we have the following stability and convergence.
Theorem 1.2.1. If 4t/x 2  1 and 4t/y 2  1, the FD scheme (1.2.2) is
stabile. Further, we have the following error estimates:

unj,k − u(xj , yk , tn ) = O(t, x 2 , y 2 ), 1  n  N. (1.2.3)

Proof. If 4t/x 2  1 and 4t/y 2  1, we have

 
n+1 2t 2t t
| uj,k |  1 − − | unj,k | + 2 (| unj+1,k | + | unj−1,k |)
x 2 y 2 x
t
+ (| unj,k+1 | + | unj,k−1 |) + t | fj,k
n
|
y 2
 un ∞ + tf ∞ , (1.2.4)

where  · ∞ is the L∞ ( ) norm. Thus, from (1.2.4), we obtain

un+1 ∞  un ∞ + tf ∞ , n = 0, 1, 2, · · · , N − 1. (1.2.5)

12 Proper Orthogonal Decomposition Methods for Partial Differential Equations

By summing (1.2.4) from 0 to n − 1, we obtain

un ∞  u0 ∞ + ntf ∞ , n = 1, 2, · · · , N. (1.2.6)

Because nt  T , from (1.2.6), we obtain

un ∞  s(x, y)∞ + T f ∞ , n = 1, 2, · · · , N, (1.2.7)

which shows that solutions of the FD scheme (1.2.2) are bounded and contin-
uously dependent on the initial value s(x, y) and source term f (x, y, t). Thus,
by Theorem 1.1.1, we deduce that the FD scheme (1.2.2) is stable. Furthermore,
the error estimates (1.2.3) are directly obtainable from approximating difference
quotients for derivatives.

Thus, as long as the time step t, the spatial steps x and y, f (x, y, t),
g(x, y, t), and s(x, y) are given, we can obtain classical FD solutions unj,k (0 
j  J, 0  k  K, 0  n  N ) for the 2D parabolic equation (1.2.1) by comput-
ing the FD scheme (1.2.2).

1.2.2 Formulation of the POD Basis

Set uni = unj,k and fin = fj,k
n (1  i  m ≡ (J + 1)(K + 1), i = kJ + k + j + 1,

0  j  J, 0  k  K, 0  n  N ). Then, the classical FD solutions for the FD

scheme (1.2.2) can be denoted by {uni }N n=1 (1  i  m). We extract the first L
sequence of solutions {uni }L n=1 (1  i  m, L N ) as snapshots. Further, we
formulate an m × L snapshot matrix
⎛ ⎞
u11 u21 · · · uL
⎜ 1 1 ⎟
⎜ u L ⎟
⎜ 2 u2 · · · u2 ⎟
2
A=⎜ . .. ⎟ . (1.2.8)
⎜ . .. .. ⎟
⎝ . . . . ⎠
u1m u2m · · · uL m

By the singular value decomposition, the snapshot matrix A has a factoriza-

tion
  
O l×(L−l)
A=U l×l VT, (1.2.9)
O (m−l)×l O (m−l)×(L−l)

where l×l = diag{σ1 , σ2 , · · · , σl } is a diagonal matrix consisting of the sin-
gular values of A according to the decreasing order σ1  σ2  · · ·  σl > 0,
U = (ϕ 1 , ϕ 2 , · · · , ϕ m ) is an m × m orthogonal matrix consisting of the orthogo-
nal eigenvectors of AAT , whereas V = (φ 1 , φ 2 , · · · , φ L ) is an L×L orthogonal
matrix consisting of the orthogonal eigenvectors of AT A, and O is a zero ma-
trix.
 Reduced-Order Extrapolation Finite Difference Schemes Chapter | 1 13

Because the number of mesh points m is much larger than that of snap-
shots L extracted, the height m of A for matrix AAT is much larger than the
width L of A for matrix AT A, but the positive eigenvalues λj (j = 1, 2, · · · , l)
of AT A and AAT are identical and satisfy λj = σj2 (j = 1, 2, · · · , l). There-
fore, we may first find the eigenvalues λ1  λ2  · · ·  λl > 0 (l = rank A) for
matrices AT A and corresponding eigenvectors ϕ j . Then, by the relationship

ϕ j = Aφ j / λj , j = 1, 2, . . . , l,

we obtain these eigenvectors ϕ j (j = 1, 2, . . . , l) corresponding to the nonzero

eigenvalues for matrix AAT .
Take
  
O M×(L−M)
AM = U M×M VT, (1.2.10)
O (m−M)×M O (m−M)×(L−M)

where M×M = diag{σ1 , σ2 , · · · , σM } is the diagonal matrix
 consisting of the
first M positive singular values of the diagonal matrix l×l . Define the norm
of matrix A as A2,2 = supu∈R L Au2 /u2 (where u2 is the l 2 norm for
vector u).
We have the following.

Lemma 1.2.2. Let  = (ϕ 1 , ϕ 2 , · · · , ϕ M ) consist of the first M eigenvectors

U = (ϕ 1 , ϕ 2 , · · · , ϕ m ). Then, we have


M
AM = σi ϕ j φ Ti = T A. (1.2.11)
i=1

Proof. According to
⎛ ⎞⎛ ⎞
σ1 ··· 0 φ T1

M
 ⎜ . ⎟⎜ ⎟
AM = σi ϕ i φ Ti = ϕ1 · · · ϕM ⎜ . .. .. ⎟ ⎜ .. ⎟,
⎝ . . . ⎠⎝ . ⎠
i=1
0 ··· σM φ TM
⎛ ⎞⎛ ⎞
σ1 ··· 0 φ T1
 ⎜ . ⎟⎜ ⎟
A= ϕ1 · · · ϕl ⎜⎝ ..
..
.
..
.
⎟ ⎜ ..
⎠⎝ .
⎟,
⎠
0 ··· σl φ Tl

we have
14 Proper Orthogonal Decomposition Methods for Partial Differential Equations

⎛ ⎞ ⎛ ⎞⎛ ⎞
ϕ T1 σ1 · · · 0 φ T1
⎜ ⎟ ⎜ . ⎟⎜ ⎟
 A =  ⎜
T . ⎟
⎝ .. ⎠ ϕ 1 · · · ϕ l ⎝ ..
⎜ ..
.
.. ⎟ ⎜ ..
. ⎠⎝ .
⎟
⎠
T
ϕM 0 ··· σl φ Tl
⎛ ⎞⎛ ⎞
σ1 · · · 0 φ T1
 ⎜ . ⎟⎜ ⎟
=  IM O ⎜ ⎝ ..
.. . ⎟⎜ . ⎟
. .. ⎠ ⎝ .. ⎠
0 · · · σl φ Tl
⎛ ⎞⎛ ⎞
σ1 · · · 0 · · · 0 φ T1
⎜ ⎟⎜ ⎟
= ⎜ . ..
⎝ .. .
..
.
. ⎟⎜ . ⎟
· · · .. ⎠ ⎝ .. ⎠
0 · · · σM · · · 0 φ Tl
⎛ ⎞⎛ ⎞
σ1 · · · 0 φ T1
 ⎜ . ⎟⎜ ⎟
= ϕ1 · · · ϕM ⎜ ⎝ ..
..
.
.. ⎟ ⎜ ..
. ⎠⎝ .
⎟ = AM .
⎠
0 · · · σM φT M

Thus, by the relationship between the matrix norm and its spectral radius,
we have

min A − B2,2 = A − AM 2,2 = A − T A2,2 = λM+1 ,
rank(B)M
(1.2.12)

where  = (ϕ 1 , ϕ 2 , · · · , ϕ M ) consists of the first M eigenvectors U = (ϕ 1 , ϕ 2 ,

· · · , ϕ m ). If the L column vectors of A are denoted by un = (un1 , un2 , · · · ,
unm )T (n = 1, 2, · · · , L), we have

un − unM 2 = (A − T A)ε n 2  A − T A2,2 ε n 2 = λM+1 ,
(1.2.13)

where unM = M j =1 (ϕ j , u )ϕ j represents the projection of u onto  =

n n

(ϕ 1 , ϕ 2 , · · · , ϕ M ), (ϕ j , u ) is the inner product of ϕ j and u , and ε n denotes

n n

the unit vector with the nth component being 1. The inequality (1.2.13) √ shows
that unM is an optimal approximation of un whose error is no more than λM+1 .
Thus,  is just an orthonormal optimal POD base of A.

1.2.3 Establishment of the POD-Based Reduced-Order Finite

Difference Scheme for the 2D Parabolic Equation
We still denote the classical FD solution vectors from the FD scheme (1.2.2) by
un = (un1 , un2 , · · · , unm )T (n = 1, 2, · · · , L, L + 1, · · · , N ). Thus, we can write
 Reduced-Order Extrapolation Finite Difference Schemes Chapter | 1 15

the FD scheme (1.2.2) in vector form as follows:

t t
un+1 = un + Bun + Cun + tF nm , (1.2.14)
x 2 y 2

where F nm = (f1n , f2n , · · · , fmn )T , and

⎛ ⎞
−1 1 0 0 ··· 0 0
⎜ ⎟
⎜ 1 −2 1 0 ··· 0 0 ⎟
⎜ ⎟
⎜ 0 1 −2 1 ··· 0 0 ⎟
⎜ ⎟
⎜ 0 0 1 −2 ··· 0 0 ⎟
B =⎜ ⎟ ,
⎜ ⎟
⎜ .. .. .. .. .. .. .. ⎟
⎜ . . . . . . . ⎟
⎜ ⎟
⎝ 0 0 0 0 ··· −2 1 ⎠
0 0 0 0 ··· 1 −1 m×m

⎛ ⎞
J times 0
⎜ −1    ⎟
⎜ ··· ··· 1 ⎟
⎜ ⎟
⎜ .. ⎟
⎜ −2 . ⎟
⎜ ⎟
⎜ .. ⎟
⎜ −2 . ⎟
⎜ ⎟
⎜ ⎟
C=⎜
⎜ 1
.. ⎟ .
⎜
. 1 ⎟
⎟
⎜ ⎟
⎜ .. ⎟
⎜ . −2 ⎟
⎜ ⎟
⎜ .. ⎟
⎜ . −2 ⎟
⎜ ⎟
⎝ 1 ·· ·  · ·· −1 ⎠
J times 0 m×m

In order to estimate a tridiagonal matrix, it is necessary to introduce the

following lemma (see [192, Theorems 1.3.1 and 1.3.2]).

Lemma 1.2.3. The tridiagonal matrix

⎛ ⎞
d b 0 0 ··· 0 0
⎜ ⎟
⎜ c a b 0 ··· 0 0 ⎟
⎜ ⎟
⎜ 0 c a b ··· 0 0 ⎟
⎜ ⎟
⎜ ··· ⎟
B̃ = ⎜ 0 0 c a 0 0 ⎟
⎜ . . . . ⎟
⎜ . . . . .. .. .. ⎟
⎜ . . . . . . . ⎟
⎜ ⎟
⎝ 0 0 0 0 ··· a b ⎠
0 0 0 0 ··· c d m×m
16 Proper Orthogonal Decomposition Methods for Partial Differential Equations

has the following eigenvalues:

√
λ̃j = a + 2 bc cos[(2j − 1)π/(2m + 1)], j = 1, 2, · · · , m.

Therefore, by using the relationship between the matrix eigenvalue and its
norm, we have

B2,2 = C2,2 =| 2 − 2 cos[(2m − 1)π/(2m + 1)] |

=| 2 − 2 cos[π − 2π/(2m + 1)] |
=| 2 + 2 cos[2π/(2m + 1)] |< 4. (1.2.15)

If we replace un of (1.2.14) with u∗n = T Aεn (n = 1, 2, · · · , L) and

u∗n = α n (n = L + 1, L + 2, · · · , N ), we obtain the PODROEFD scheme as
follows:
⎧
⎪
⎪ u∗n = T Aεn , n = 1, 2, · · · , L,
⎪
⎨
t t
⎪ α n+1 = α n + Bα n + Cα n + tF nm , (1.2.16)
⎪
⎪ x 2 y 2
⎩
n = L, L + 1, · · · , N − 1,

where α n = (α1n , α2n , · · · , αM

n )T are vectors yet to be determined.

By using the orthogonal vectors in T multiplied by Eq. (1.2.16), we obtain

⎧
⎪
⎪
⎨ α =  u , n = 1, 2, · · · , L,
n T n

⎪ α n+1 = α n + x t
2  Bα + y 2  Cα + t F m ,
T n t T n T n (1.2.17)
⎪
⎩
n = L, L + 1, · · · , N − 1.

After α n (n = L, L+1, · · · , N ) are obtained from the system of Eq. (1.2.17),

we can obtain the PODROEFD solution vectors for Eq. (1.2.1) as follows:

u∗n = α n , n = 1, 2, · · · , L, L + 1, · · · , N. (1.2.18)

Further, we can obtain the PODROEFD solution components for Eq. (1.2.1)
as follows:

u∗n ∗n
j,k = ui , 0  j  J, 0  k  K,
1  i = k(J + 1) + j + 1  m = (K + 1)(J + 1). (1.2.19)

Remark 1.2.1. It is easily seen that the classical FD scheme (1.2.2) at each
time level contains m unknown quantities, whereas the system of Eqs. (1.2.17)–
(1.2.18) at the same time level (when n > L) contains only M unknown quanti-
ties (M L m, usually M = 6, but m = O(104 ) ∼ O(106 )). Therefore, the
PODROEFD model (1.2.17)–(1.2.18) includes very few degrees of freedom and
does not involve repeated computations. Here, we extract the snapshots from the
 Reduced-Order Extrapolation Finite Difference Schemes Chapter | 1 17

first L classical FD solutions; but when we solve real-world problems we may,

instead, extract snapshots from the samples of experiments of physical system
trajectories.

1.2.4 Error Estimates of the Reduced-Order Finite Difference

Solutions for the 2D Parabolic Equation
First, from (1.2.13), we obtain

un − u∗n 2 = un − unM 2 = (A − T A)ε n 2  λM+1 ,
n = 1, 2, · · · , L. (1.2.20)

Rewrite the second equation of the system of Eqs. (1.2.16) as follows:

t t
u∗n+1 = u∗n + 2
Bu∗n + Cu∗n + tF nm ,
x y 2
n = L, L + 1, · · · , N − 1. (1.2.21)

Put δ = tB2,2 /x 2 + tC2,2 /y 2 . By subtracting (1.2.21) from

(1.2.14) and taking norm, from (1.2.20), we have

un+1 − u∗n+1 2  (1 + δ)un − u∗n 2

 ···
 (1 + δ)n+1−L uL − u∗L 2

 (1 + δ)n+1−L λM+1 , n = L, L + 1, · · · , N − 1. (1.2.22)

By summarizing the above discussion and noting that the absolute value of
each vector component is less than the vector norm, from (1.2.3), we obtain the
following theorem.

Theorem 1.2.4. The errors between the solutions unjk from the classical FD
scheme and the solutions u∗n j k from the PODROEFD scheme (1.2.17)–(1.2.18)
satisfy the following estimates:

| unjk − u∗n
j k | E(n) λM+1 , 1  j  J, 1  k  K, 1  n  N, (1.2.23)

where E(n) = 1 (1  n  L), whereas E(n) = (1 + δ)n−L (L + 1  n  N ),

δ = tB2,2 /x 2 + tC2,2 /y 2 . Further, the errors between the accurate
solution u(x, y, t) from Eq. (1.2.1) and the solutions u∗n
j k from the POD-based
reduced-order FD scheme (1.2.17)–(1.2.18) satisfy the following estimates:

| u(xj , yk , tn ) − u∗n
j k |= O(E(n) λM+1 , t, x , y ),
2 2

1  j  J, 1  k  K, 1  n  N. (1.2.24)
18 Proper Orthogonal Decomposition Methods for Partial Differential Equations

√
Remark 1.2.2. The error terms containing λM+1 in Theorem 1.2.4 are
caused by the POD-based reduced-order for the classical FD scheme, which
could be used to select
√ the number of the POD basis, i.e., it is necessary to
take M such that λM+1 = O(t, x 2 , y 2 ). Whereas E(n) = (1 + δ)n−L
(L + 1  n  N ) are caused by extrapolating iterations, which√could be
used as the guide for renewing the POD basis, i.e., if (1 + δ)n−L λM+1 >
max(t, x 2 , y 2 ), it is √
necessary to update the POD basis. If we take λM+1
that satisfies (1 + δ)N−L λM+1 = O(t, x 2 , y 2 ), then the PODROEFD
scheme (1.2.17)–(1.2.18) is convergent and, thus, we do not have to update the
POD basis.

1.2.5 The Implementation of the Algorithm of the POD-Based

Reduced-Order Finite Difference Scheme for the 2D
Parabolic Equation
In order to facilitate the use of the PODROEFD scheme for the 2D parabolic
equation, the following implementation steps of the algorithm for the PO-
DROEFD scheme (1.2.17)–(1.2.18) are helpful.
Step 1. Classical FD computation and extraction of snapshots
Write the classical FD scheme (1.2.2) in vector form (1.2.14) and find the so-
lution vectors un = (un1 , un2 , · · · , unm )T (n = 1, 2, · · · , L) of (1.2.14) at the first
few L steps (in the following, say, take L = 20 in Section 1.2.6).
Step 2. Snapshot matrix A and eigenvalues of AT A
Formulate snapshot matrices A = (uni )m×L and compute the eigenvalues λ1 
λ2  · · ·  λl > 0 (l = rankA) and the eigenvectors φ j (j = 1, 2, · · · , M̃) of
matrices AT A.
Step 3. Choice of POD basis
For the error tolerance μ = O(t, x 2 , y 2 ), decide the numbers M (M 
M̃) of POD bases such that λu(M+1)  μ and formulate the POD bases  =

(ϕ 1 , ϕ 2 , · · · , ϕ M ) (where ϕ j = Aφ j / λj , j = 1, 2, · · · , M).
Step 4. Solve/compute the PODROEFD model
Solve the PODROEFD scheme (1.2.17)–(1.2.18) to obtain the reduced-order so-
lution vectors u∗n = (u∗n ∗n ∗n T
1 , u2 , · · · , um ) ; further, obtain the component forms
∗n ∗n
uj,k = ui (0  j  J , 0  k  K, i = k(J + 1) + j + 1, 1  i  m =
(K + 1)(J + 1)).
Step 5. Check accuracy and renew POD basis to continue

Set δ = tB2,2 /x 2 + tC2,2 /y 2 . If (1 + δ)n−L λu(M+1)  μ. Then
u∗n = (u∗n ∗n ∗n T
1 , u2 , · · · , um ) is just the solution vectors for the PODROEFD
scheme(1.2.17)–(1.2.18) that satisfy the accuracy requirement. Else, i.e., if (1 +
δ)n−L λu(M+1) > μ, put u1 = u∗(n−L) , u2 = u∗(n−L+1) , . . . , uL = u∗(n−1) and
return to Step 2.
Another random document with
no related content on Scribd:
 CHAPTER III
CRUSTACEA (CONTINUED): COPEPODA

Order II. Copepoda.

The Copepods are small Crustacea, composed typically of about sixteen

segments, in which the biramous type of limb predominates. They are
devoid of a carapace. Development proceeds gradually by the addition
posteriorly of segments to a Nauplius larval form. Paired compound eyes
are absent, except in Branchiura, the adult retaining the simple eye of the
Nauplius.
In a typical Copepod, such as Calanus hyperboreus (Fig. 25), we can
distinguish the following segments with their appendages: a
cephalothorax, carrying a pair of uniramous first antennae (1st Ant.); a
pair of biramous second antennae (2nd Ant.); mandibles (Md.) with biting
gnathobases and a palp, and a pair of foliaceous first maxillae (Mx.1). Two
pairs of appendages follow, which were looked upon as the two branches
of the second maxillae, but it is now certain that they represent two pairs
of appendages, which may be called second maxillae (Mx.2) and
maxillipedes (Mxp.) respectively. Behind these are five pairs of biramous
swimming feet, the first pair (Th.1) attached to the cephalothorax, the
succeeding four pairs to four distinct thoracic somites. Behind the thorax
is a clearly delimited abdomen composed of five segments, the first of
which (Abd.1) carries the genital opening, and the last a caudal furca.
The Copepods exhibit a great variety of structure, and their
classification is attended with great difficulties. Claus[36] based his
attempt at a natural classification on the character of the mouth and its
appendages, dividing the free-living and semi-parasitic forms as
Gnathostomata from the true parasites or Siphonostomata. This division,
although convenient, breaks down in many places, and it is clear that the
parasitic mode of life has been acquired more than once in the history of
Copepod evolution, while the free-living groups do not constitute a
natural assemblage.
Fig. 25.—Calanus hyperboreus, × 30. Abd1, First abdominal segment;
1st Ant, 2nd Ant, 1st and 2nd antennae; Md, mandible; Mx1, Mx2, 1st
 and 2nd maxillae; Mxp, maxillipede; Th1, 1st thoracic appendage.
(After Giesbrecht.)

Giesbrecht has more recently[37] founded a classification of the free-

living pelagic Copepods upon the segmentation of the body and certain
secondary sexual characters, and he has hinted[38] that this scheme of
classification applies to the semi-parasitic and parasitic forms. Although
much detail remains to be worked out and the position of some families is
doubtful, Giesbrecht’s scheme is the most satisfactory that has hitherto
been suggested, and will be adopted in this chapter.
The peculiarity in structure of the Argulidae, a small group of
ectoparasites on fresh water fish, necessitates their separation from the
rest of the Copepods (Eucopepoda) as a separate Branch, Branchiura.

BRANCH I. EUCOPEPODA.

Sub-Order 1. Gymnoplea.

The division between the front and hind part of the body falls
immediately in front of the genital openings and behind the fifth thoracic
feet. The latter in the male are modified into an asymmetrical copulatory
organ.

TRIBE I. AMPHASCANDRIA.

The first antennae of the male are symmetrical, with highly-developed

sensory hairs.
Fam. Calanidae.—The Calanidae are exclusively marine Crustacea,
and form a common feature of the pelagic plankton in all parts of the
world. Some species of the genus Calanus often occur in vast shoals,
making the sea appear blood-red, and they furnish a most important
article of fish food. These swarms appear to consist chiefly of females, the
males being taken rarely, and only at certain seasons of the year. Some of
the Calanidae are animals of delicate and curious form, owing to the
development of plumed iridescent hairs from various parts of their body,
which may often exhibit a marked asymmetry, as in the species figured,
Calocalanus plumulosus (Fig. 26), from the Mediterranean.
 Sars makes a curious
[39]
observation with regard to the
distribution of certain Calanidae.
He reports that along the whole
route of the “Fram,” species such
as Calanus hyperboreus and
Euchaeta norwegica were taken
at the surface, which, in the
Norwegian fjords, only occur at
depths of over 100 fathoms. He
suggests that the Norwegian
individuals, instead of migrating
northwards as the warmer
climate supervened, have sought
boreal conditions of temperature
by sinking into the deeper waters.

Fig. 26.—Calocalanus plumulosus, ×

15. (After Giesbrecht.)

TRIBE II.
HETERARTHRANDRIA.

The first antennae of the male are asymmetrical, one, usually the
right, being used as a clasping organ.
The males of the Centropagidae, Candacidae and Pontellidae,
besides possessing the asymmetrically modified thoracic limbs of the
fifth pair also exhibit a modification of one of the first antennae,
which is generally thickened in the middle, and has a peculiar joint in
it, or geniculation, which enables it to be flexed and so used as a
clasping organ for holding the female.
Fam. 1.—Centropagidae.—These Copepods are very common
in the pelagic plankton, and some of the species vie with the
Calanidae in plumed ornaments, e.g. Augaptilus filigerus, figured by
Giesbrecht in his monograph. The use of these ornaments, which are
possessed by so many pelagic Copepods, is entirely obscure.[40]
Certain of the Centropagidae live in fresh water. Thus Diaptomus is
an exclusively fresh-water genus, and forms a most important
constituent of lake-plankton; various species of Heterocope occur in
the great continental lakes, and certain Eurytemora go up the
estuaries of rivers into brackish water.
An excellent work on the fresh-water Copepods of Germany has
been written by Schmeil,[41] who gives analytical tables for
distinguishing various genera and species. The three fresh-water
families are the Centropagidae, Cyclopidae, and Harpacticidae (see
p. 62). The Centropagidae may be sharply distinguished from the
other fresh-water families by the following characters:—The
cephalothorax is distinctly separated from the abdomen; the first
antennae are long and composed of 24–25 segments, in the male
only a single antenna (generally the right) being geniculated and
used as a clasping organ. The fifth pair of limbs are not rudimentary;
a heart is present, and only one egg-sac is found in the female. The
second antennae are distinctly biramous.
Diaptomus.—The furcal processes are short, at most three times as
long as broad; endopodite of the first swimming appendage 2–
jointed, endopodites of succeeding legs 3–jointed.
Heterocope.—The furcal processes are short, at most twice as long
as broad; endopodites of all swimming legs 1–jointed.
Eurytemora.—The furcal processes are long, at least three and a
half times as long as broad; the endopodite of the first pair of legs
1–jointed, those of the other pairs 2–jointed.
It has been known for a long time that some of the marine
Copepods are phosphorescent, and, indeed, owing to their numbers
in the plankton, contribute very largely to bring about that liquid
illumination which will always excite the admiration of seafarers. In
northern seas the chief phosphorescent Copepods belong to
Metridia, a genus of the Centropagidae; but in the Bay of Naples
Giesbrecht[42] states that the phosphorescent species are the
following Centropagids: Pleuromma abdominale and P. gracile,
Leuckartia flavicornis and Heterochaeta papilligera; Oncaea
conifera is also phosphorescent. It is often stated that Sapphirina (p.
69) is phosphorescent, but its wonderful iridescent blue colour is
purely due to interference
colours, and has nothing to do
with phosphorescence.
Giesbrecht has observed that the
phosphorescence is due to a
substance secreted in special
skin-glands, which is jerked into
the water, and on coming into
contact with it emits a
phosphorescent glow. This
substance can be dried up
completely in a desiccated
specimen and yet preserve its
phosphorescent properties, the
essential condition for the actual
emission of light being contact
with water. Similarly, specimens
preserved in glycerine for a long
period will phosphoresce when
compressed in distilled water.
From this last experiment
Giesbrecht concludes that the
phosphorescence can hardly be
due to an oxidation process, but
the nature of the chemical
reaction remains obscure.
Fam. 2. Candacidae.—This
family comprises the single genus
Candace, with numerous species
distributed in the plankton of all
seas. Some species, e.g. C.
pectinata, Brady, have a
practically world-wide
distribution, this species being
recorded from the Shetlands and
from the Philippines.
Fam. 3. Pontellidae.—This is
a larger family also comprising
 widely distributed species found
Fig. 27.—Dorsal view of Anomalocera
pattersoni, ♂, × 20. (After Sars.)in the marine plankton.
Anomalocera pattersoni (Fig. 27)
is one of the commonest elements in the plankton of the North Sea.

Sub-Order 2. Podoplea.

The boundary between the fore and hind part of the body falls in
front of the fifth thoracic segment. The appendages of the fifth
thoracic pair in the male are never modified as copulatory organs.

TRIBE I. AMPHARTHRANDRIA.

The first antennae in the male differ greatly from those in the
female, being often geniculated and acting as prehensile organs.
Fams. 1–2. Cyclopidae and Harpacticidae, and other allied
families, are purely free-living forms; they are not usually pelagic in
habit, but prefer creeping among algae in the littoral zone or on the
sea-bottom, or especially in tidal pools. Some genera are,
nevertheless, pelagic; e.g. Oithona among Cyclopidae; Setella,
Clytemnestra, and Aegisthus among Harpacticidae.
The sketch (Fig. 28) of Euterpe acutifrons ♀ , a species widely
distributed in the Mediterranean and northern seas, exhibits the
structure of a typical Harpacticid, while Fig. 29 shows the form of the
first antenna in the male.
Several fresh-water representatives of these free-living families
occur. The genus Cyclops (Cyclopidae) is exclusively fresh-water,
while many Harpacticidae go up into brackish waters: for example
on the Norfolk Broads, Mr. Robert Gurney has taken Tachidius
brevicornis, Müller, and T. littoralis, Poppe; Ophiocamptus
brevipes, Sars; Mesochra lilljeborgi, Boeck; Laophonte littorale, T.
and A. Scott; L. mohammed, Blanchard and Richard; and
Dactylopus tisboides, Claus.
Schmeil[43] gives the following scheme for identifying the fresh-
water Cyclopidae and Harpacticidae (see diagnosis of Centropagidae
on p. 59):—
 Fam. 1. Cyclopidae.—The
cephalothorax is clearly separated
from the abdomen. The first
antennae of the female when bent
back do not stretch beyond the
cephalothorax; in the male both
of them are clasping organs. The
second antennae are without an
exopodite. The fifth pair of limbs
are rudimentary, there is no
heart, and the female carries two
egg-sacs.
Cyclops.—Numerous species, split
up according to segmentation of
rudimentary fifth pair of legs,
number of joints in antennae, etc.
Fam. 2. Harpacticidae.—
The cephalothorax is not clearly
separated from the abdomen. The
first antennae are short in both
sexes, both being clasping organs
in the male. The second antennae
have a rudimentary exopodite.
The fifth pair of limbs are
rudimentary and plate-shaped; a
heart is absent, and the egg-sacs
Fig. 28.—Euterpe acutifrons, ♀ , × 70. of the female may be one or two
Abd.1, 1st abdominal segment; Th.5, in number.
5th thoracic segment. (After
Giesbrecht.) 1. Ophiocamptus (Moraria).—Body
worm-shaped; first antennae of
female 7–jointed, rostrum
forming a broad plate.
2. Body not worm-shaped; first antennae of female 8–jointed,
rostrum short and sharp.
(a) Endopodites of all thoracic limbs 3–jointed. The first
antennae in female distinctly bent after the second joint.
Nitocra.
 (b) Endopodite of at least
the fourth limb 2–jointed;
first antennae in female
not bent. Canthocamptus.
3. Ectinosoma.—Body as in
2, but first antennae are very
short, and the maxillipede
does not carry a terminal
hooked seta as in 1 and 2.
Fam. 3. Peltiidae.[44]—This is
an interesting family, allied to the
Harpacticidae, and includes
species with flattened bodies
somewhat resembling Isopods,
and a similar habit of rolling
themselves up into balls. No
parasitic forms are known,
though Sunaristes paguri on the
French and Scottish coasts is said
to live commensally with hermit-
crabs.

Fig. 29.—First antenna of Euterpe

We have now enumerated the acutifrons, ♂. (After Giesbrecht.)
chief families of free-living
Copepods; the rest are either true
parasites or else spend a part of their lives as such. A number of the
semi-parasitic and parasitic Copepods can be placed in the tribe
Ampharthrandria owing to the characters of their antennae; but it
must be remembered that many parasitic forms have given up using
the antennae as clasping organs; however, the sexual differences in
the antennae, and the fact that many of the species which have lost
the prehensile antennae in the male have near relations which
preserve it, enable us to proceed with some certainty. The adoption
of this classification necessitates our separating many families which
superficially may seem to resemble one another, e.g. the semi-
parasitic families Lichomolgidae and Ascidicolidae, and the
Dichelestiidae from the other fish-parasites; it also necessitates our
treating the presence of a sucking mouth as of secondary importance.
This characteristic must certainly, however, have been acquired more
than once in the history of the Copepods, for instance in the
Asterocheridae and in the fish-parasites, while it sometimes happens
that genera belonging to a typically Siphonostomatous group possess
a gnathostome, or biting mouth, e.g. Ratania among the
Asterocheridae. Again, it is impossible even if we use the character of
the mouth as a criterion to place together all the true parasites on
fishes in one natural group, because the Bomolochidae and
Chondracanthidae, which are otherwise closely similar to the rest of
the fish-parasites, possess no siphon. It seems plain, therefore, that
the parasitic habit has been acquired several times separately by
diverging stocks of free-swimming Copepods, and that it has resulted
in the formation of convergent structures.
Fam. 4. Monstrillidae.[45]—These are closely related to the
Harpacticidae. The members of this curious family are parasitic
during larval life and actively free-swimming when adult. There are
three genera, Monstrilla, Haemocera, and Thaumaleus. The best
known type is Haemocera danae (often described as Monstrilla
danae). In the adult state (Fig. 30) there are no mouth-parts; the
mouth is exceedingly small and leads into a very small stomach,
which ends blindly, while the whole body contains reserve food-
material in the form of brown oil-drops. The sole appendages on the
head are the first antennae; but on the thorax biramous feet are
present by means of which the animal can swim with great rapidity.
This anomalous organisation receives an explanation from the
remarkable development through which the larva passes. The larva is
liberated from the parent as a Nauplius with the structure shown in
Fig. 31; it does not possess an alimentary canal. It makes its way to a
specimen of the Serpulid worm, Salmacina dysteri, into the
epidermis of which it penetrates by movements of the antennae,
hanging on all the time by means of the hooks on the mandibles.
From the epidermis it passes through the muscles into the coelom of
the worm, and thence into the blood-vessels, usually coming to rest
in the ventral blood-vessel. As the Nauplius migrates, apparently by
amoeboid movements of the whole body, it loses all its appendages,
the eye degenerates, and the body is reduced to a minute ovoid mass
of cells, representing ectoderm and endo-mesoderm, surrounded by
a chitinous membrane (Fig. 32, A). Arrived in the ventral blood-
 vessel it begins to grow, and the
first organ formed is a pair of
fleshy outgrowths representing
the second antennae (Fig. 32, B),
which act as a nutrient organ
intermediary between host and
parasite. The adult organs now
begin to be differentiated, as
shown in Fig. 32, C, from the
undifferentiated cellular elements
of the Nauplius, the future adult
organism being enclosed in a
spiny coat from which it escapes.
At this stage it occupies a large
part of its host’s body, lying in the
distended ventral blood-vessel,
and it escapes to the outside
world by rupturing the body-wall
of the worm, leaving behind it the
second antennae, which have
Fig. 30.—Haemocera danae, × 40. A, performed their function as a
Side view ♀ ; B, ventral view ♂ . Ant.1, kind of placenta. Malaquin, to
1st antenna; e, eye; ov, ovary; ovd, whom we owe this account,
oviduct; St, stomach; Th.1, 1st thoracic
appendage; Th.5, 5th thoracic segment;
makes the remarkable statement
vd, vas deferens. (After Malaquin.) that if two or three Monstrillid
Nauplii develop together in the
same host they are always males,
if only one it may be either male or female. The only parallel to this
extraordinary life-history is found in the Rhizocephala (see pp. 96–
99).
Fig. 31.—Free-swimming Nauplius
larva of Haemocera danae; Ant.1,
Ant.2, 1st and 2nd antennae; e, remains
of eye; Md, mandible. (After Malaquin.)
Fig. 32.—Later stages in the development of Haemocera danae.
Abd, Abdomen; Ant.1, Ant.2, 1st and 2nd antennae; ch, chitinous
investment; e, eye; Ect, ectoderm; En, endoderm; Mes,
mesoderm; Mes & en, mesoderm and endoderm; R, rostrum; St,
mouth and stomach; Th, thoracic appendages. (After Malaquin.)
 Fam. 5. Ascidicolidae.[46]—
Although the members of this
family, which live
semiparasitically in the branchial
sac or the gut of Ascidians, betray
their Ampharthrandrian nature
by the sexual differences of their
first antennae, only two genera,
Notodelphys and Agnathaner,
possess true prehensile antennae.
According as the parasitism is
Fig. 33.—Side view of Doropygus
pulex, ♀ , × 106. Abd.1, 1st abdominal more or less complete, the buccal
segment; Ant.1, 1st antenna; b.p, appendages either retain their
brood-pouch; Th.1, 1st thoracic masticatory structure or else
appendage; Th.4, 4th thoracic segment. become reduced to mere organs
(After Canu.) of fixation. In Notodelphys both
sexes can swim actively and
retain normal mouth-parts; they live parasitically, or perhaps
commensally, in the branchial cavities of Simple or Compound
Ascidians, feeding on the particles swept into the respiratory
chamber of the host. They leave their host at will in search of a new
home, and are frequently taken in the plankton.
Doropygus (Fig. 33), a genus widely distributed in the North Sea
and Mediterranean, also inhabiting the branchial sac of Ascidians, is
more completely parasitic, and the female cannot swim actively.
Forms still more degraded by a parasitic habit are Ascidicola rosea
(especially abundant in the stomach of Ascidiella scabra at
Concarneau), in which the female has lost its segmentation, the
mouth-parts and thoracic legs being purely prehensile, and various
species of Enterocola, parasitic in the stomach of Compound
Ascidians, in which the female is a mere sac incapable of free motion,
while the male preserves its swimming powers and a general
Cyclops-form (Fig. 34). We have here the first instance of the
remarkable parallelism between the degree of parasitism and the
degree of sexual dimorphism, a parallelism which holds with great
regularity among the Copepoda, and can be also extended to other
classes of parasitic animals.
 Fam. 6. Asterocheridae.[47]
—These forms retain the power of
swimming actively, and are very
little modified in outward
appearance by their parasitic
mode of life (Fig. 35), though they
possess a true siphon in which the
styliform mandibles work. The
siphon is formed by the upper
and lower lips, which are
produced into a tube with three
longitudinal ridges; in the outer
grooves are the mandibles, while
the inner groove forms the
sucking siphon (see transverse
section, Fig. 36). In Ratania,
however, there is no siphon. The
Fig. 34.—Enterocola fulgens. A, first antennae possess a great
Ventral view of ♀, × 35; B, side view of number of joints, and may be
♂ , × 106. Abd.1, 1st abdominal geniculated in the male
segment; Ant.1, Ant.2, 1st and 2nd (Cancerilla). The members of this
antennae; c.m, gland-cells; n, ventral family live as ectoparasites on
nerve-cord; og, oviducal gland; ov, various species of Echinoderms,
ovary; po, vagina; Th.1, 1st thoracic Sponges, and Ascidians, but they
appendage; Th.4, Th.5, 4th and 5th
thoracic segments. (After Canu.) frequently change their hosts, and
it appears that one and the same
species may indifferently suck the
juices of very various animals, and even of Algae. Cancerilla
tubulata, however, appears to live only on the Brittle Starfish,
Amphiura squamata.
Fam. 7. Dichelestiidae.—The males and females are similarly
parasitic, and the body in both is highly deformed, the segmentation
being suppressed and the thoracic limbs being produced into
formless fleshy lobes; they are placed among the Ampharthrandria
owing to sexual differences in the form of the first antennae. There is
a well-developed siphon in which the mandibular stylets work,
except in Lamproglena, parasitic on the gills of Cyprinoid fishes; the
succeeding mouth-parts are prehensile.
 The majority of the species are
parasitic on the gills of various
fish (Dichelestium on the
[48]
Sturgeon, Lernanthropus on
Labrax lupus, Serranus scriba,
etc.), but Steuer[49] has recently
described a Dichelestiid
(Mytilicola) from the gut of
Mytilus galloprovincialis off
Trieste. This animal and
Lernanthropus are unique among
Crustacea through the possession
of a completely closed blood-
vascular system which contains a
red fluid; the older observers
believed this fluid to contain
haemoglobin, but Steuer, as the
result of careful analysis, denies
this. The parasite on the gills of
the Lobster, Nicothoe astaci,
possibly belongs here.
The inclusion of Nicothoe and
the Dichelestiidae among the Fig. 35.—Asterocheres violaceus, ♀ ,
Ampharthrandria rests on a with egg-sacs, × 57. (After Giesbrecht.)
somewhat slender basis; this
basis is afforded by the fact that
none of the parasitic Isokerandria have more than seven joints in the
first antennae, whereas Nicothoe and some of the Dichelestiidae[50]
have more numerous joints. In most of the Dichelestiidae, however,
the number of joints is less than seven and practically equal in the
two sexes.
Fig. 36.—Diagrammatic transverse
section through the distal part of the
siphon of Rhynchomyzon
purpurocinctum (Asterocheridae). Md,
mandible. (After Giesbrecht.)

TRIBE II. ISOKERANDRIA.

The first antennae are short, similar in the two sexes, and are
never used by the male as clasping organs. This function may be
subserved by the second maxillae.
Fams. Oncaeidae, Corycaeidae, Lichomolgidae, Ergasilidae,
Bomolochidae, Chondracanthidae, Philichthyidae,
Nereicolidae, Hersiliidae, Caligidae, Lernaeidae,
Lernaeopodidae, Choniostomatidae.
The families Oncaeidae and Corycaeidae contain pelagic forms of
flattened shape and great swimming powers, but the structure of the
mouth-parts in the Corycaeidae points to a semi-parasitic habit.
Fam. 1. Oncaeidae.—This family, including the genera Oncaea,
Pachysoma, etc., does not possess the elaborate eyes of the next
family, nor is the sexual dimorphism so marked.
Fam. 2. Corycaeidae.—These are distinguished from the
Oncaeidae, not only by their greater beauty, but also by the
possession of very elaborate eyes, which are furnished with two
lenses, one at each end of a fairly long tube. The females of
Sapphirina are occasionally found in the branchial cavity of Salps,
and their alimentary canal never contains solid particles, but is filled
with a fluid substance perhaps derived by suction from their prey. S.
opalina may occur in large shoals, when the wonderful iridescent
blue colour of the males makes the water sparkle as it were with a
sort of diurnal phosphorescence. The animal, however, despite the
opinion of the older observers, is not truly phosphorescent. It may be
that the ornamental nature of some of the males is correlated with
the presence of the curious visual organs, which are on the whole
better developed in the females than in the males. As in so many
pelagic Copepods, the body and limbs may bear plumed setae of
great elaboration and beautiful colour, e.g. Copilia vitrea (Fig. 37).
We now pass on to the rest of the parasitic Copepods,[51] which
probably belong to the tribe Isokerandria, and we meet with the
same variety of degrees of parasitism as in the Ampharthrandria,
often leading to very similar results.
 Fig. 37.—Copilia vitrea (Corycaeidae), ♀, × 20. (After Giesbrecht.)

In the first seven families mentioned below there is no siphon. The

Lichomolgidae and Ergasilidae have not much departed from the
free-living forms just considered, retaining their segmentation,
though in the Ergasilidae the body may be somewhat distorted (Fig.
39). In both families the thoracic swimming feet are of normal
constitution.
Fam. 3. Lichomolgidae.[52]
—These are semi-parasitic in a
number of animals living on the
sea-bottom, such as Actinians,
Echinoderms, Annelids, Molluscs,
and Tunicates. Lichomolgus
agilis (Fig. 38) occurs in the
North Sea, Atlantic, and
Mediterranean, on the gills of
large species of the Nudibranch,
Doris, while L. albeus is found in
the peribranchial cavity and
cloaca of various Ascidians.
Sabelliphilus may infect the gills
of Annelids such as Sabella, and
is common at Liverpool.
Fam. 4. Ergasilidae.—
Thersites (Fig. 39) is parasitic on
the gills of various fishes, e.g. T.
gasterostei, which is common on
Gasterosteus aculeatus on the
French and North Sea coasts, and
may even be found on specimens
of the fish that have run up the
River Forth into fresh water. The
animal possesses claw-like
second antennae by which it
clings to its host.
Fig. 38.—Lichomolgus agilis, × 10.
Abd. 1, 1st abdominal segment; cpth,
cephalothorax; Th.1, 1st thoracic
segment; Th.5, 5th thoracic appendage.
(After Canu.)